goblin-agents 3.1.0 → 3.2.0

package/lib/llm/aiAgent.js

@@ -13,7 +13,12 @@ const {
 } = require('xcraft-core-stones');
 const {getMutex} = require('xcraft-core-utils/lib/locks.js');
 const LLMProviders = require('./providers.js');
-const {OCR_SYSTEM_PROMPT, getPDFImages, splitThinkTags} = require('./utils.js');
+const {
+  OCR_SYSTEM_PROMPT,
+  getPDFImages,
+  splitThinkTags,
+  getContextUsage,
+} = require('./utils.js');
 const Logger = require('../logger.js');
 
 const CONVERSATION_TITLE_PROMPT = `Create a concise and informative title that summarizes the main topic of the conversation.
@@ -30,6 +35,15 @@ class MessageShape {
   tool_name = option(string);
 }
 
+class UsageShape {
+  used = number;
+  reserved = number;
+  remaining = number;
+  total = number;
+  percentage = number;
+  nearLimit = boolean;
+}
+
 class OptionsShape {
   // Active NUMA (Non-Uniform Memory Access) pour optimiser la gestion mémoire sur les architectures multi-sockets.
   numa = option(boolean);
@@ -150,6 +164,7 @@ class AiAgentShape {
   role = string; //agent role
   prompt = string; //default system prompt
   contexts = record(string, string); //additionnal system prompt context
+  usages = record(string, UsageShape);
   reasonings = record(
     string,
     enumeration(false, true, 'low', 'medium', 'high')
@@ -204,6 +219,7 @@ class AiAgentLogic extends Elf.Archetype {
     provider: defaultValues.provider,
     model: defaultValues.model,
     otherModels: defaultValues.otherModels,
+    usages: {},
     messages: {},
     contexts: {},
     reasonings: {},
@@ -239,10 +255,21 @@ class AiAgentLogic extends Elf.Archetype {
     this.patch(patch);
     state.version = version;
   }
+  _updateContextUsage(contextId, usage) {
+    this.state.usages[contextId] = usage;
+  }
 
   reset(contextId) {
     const {state} = this;
     state.messages[contextId] = [];
+    state.usages[contextId] = {
+      used: 0,
+      reserved: 0,
+      remaining: this.state.options?.num_ctx || 4096,
+      total: this.state.options?.num_ctx || 4096,
+      percentage: 0,
+      nearLimit: false,
+    };
     if (state.contexts[contextId]) {
       delete state.contexts[contextId];
     }
@@ -509,7 +536,7 @@ class AiAgent extends Elf {
           ],
         },
       ];
-      const result = await this._provider.chat({
+      const {result} = await this._provider.chat({
         model: this.state.model,
         reasoning: this.state.reasoning,
         messages,
@@ -542,7 +569,7 @@ class AiAgent extends Elf {
         ],
       },
     ];
-    const result = await this._provider.chat({
+    const {result} = await this._provider.chat({
       model: this.state.model,
       reasoning: this.state.reasoning,
       messages,
@@ -594,6 +621,37 @@ class AiAgent extends Elf {
     this.logic.set(contextId, messages);
   }
 
+  async getContextUsage(contextId) {
+    const usage = this.state.usages[contextId];
+    if (!usage) {
+      return {
+        used: 0,
+        reserved: 0,
+        remaining: this.state.options?.num_ctx || 4096,
+        total: this.state.options?.num_ctx || 4096,
+        percentage: 0,
+        nearLimit: false,
+      };
+    }
+    return {
+      used: usage.used,
+      reserved: usage.reserved,
+      remaining: usage.remaining,
+      total: usage.total,
+      percentage: usage.percentage,
+      nearLimit: usage.nearLimit,
+    };
+  }
+
+  async _updateContextUsage(contextId, response) {
+    const usage = getContextUsage(
+      response,
+      this.state.options?.num_ctx || undefined,
+      this.state.options?.num_predict || undefined
+    );
+    this.logic._updateContextUsage(contextId, usage);
+  }
+
   /**
    * Set chat session context, model and level of reasoning
    * @param {string} contextId
@@ -623,7 +681,7 @@ class AiAgent extends Elf {
     return {context, model, reasoning};
   }
 
-  async chat(contextId, question, userDesktopId) {
+  async chat(contextId, question, userDesktopId, disableTool = false) {
     const logger = Logger.get(contextId, this);
 
     await this._mutex.lock(this.id);
@@ -670,11 +728,11 @@ ${context}
       messages,
       options,
       format: this.state.format?.toJS(),
-      tools: this.state.tools?.toJS(),
+      tools: disableTool ? undefined : this.state.tools?.toJS(),
     };
 
-    const result = await this._provider.chat(chatCompletionRequest);
-
+    const {result, usage} = await this._provider.chat(chatCompletionRequest);
+    await this._updateContextUsage(contextId, usage);
     logger.push('aiAgent.chat', {
       id: this.id,
       model: chatCompletionRequest.model,
@@ -684,10 +742,10 @@ ${context}
       ),
       question,
       reasoning: result.message?.reasoning,
-      callTools: !!result.message.tool_calls?.length,
+      callTools: disableTool ? 'disabled' : !!result.message.tool_calls?.length,
     });
 
-    if (this.state.toolServiceId && this.state.tools) {
+    if (this.state.toolServiceId && this.state.tools && !disableTool) {
       const used = await this.callTools(
         contextId,
         result.message,
@@ -695,7 +753,7 @@ ${context}
       );
       if (used) {
         //leave for a second turn
-        this.log.dbg(`🐈 Chat session ended : ${contextId}`);
+        this.log.dbg(`🐈 Chat session ended by a tool call: ${contextId}`);
         return true;
       }
     }

package/lib/llm/providers/ollama.js

@@ -37,9 +37,9 @@ class OllamaProvider extends LLMProvider {
     });
     if (useFormat) {
       chat.message.content = JSON.parse(chat.message.content);
-      return chat;
+      return {result: chat};
     } else {
-      return chat;
+      return {result: chat};
     }
   }
 

package/lib/llm/providers/opanAi.js

@@ -53,7 +53,7 @@ class OpenAIProvider extends LLMProvider {
         methods: ['POST'],
       },
     });
-    const {choices} = chat;
+    const {choices, usage} = chat;
     if (chat.error) {
       const err = new Error(
         chat.error?.metadata
@@ -67,9 +67,9 @@ class OpenAIProvider extends LLMProvider {
       choices[0].message.content = JSON.parse(
         choices[0].message.content
       ).results;
-      return choices[0];
+      return {result: choices[0], usage};
     } else {
-      return choices[0];
+      return {result: choices[0], usage};
     }
   }
 
@@ -96,8 +96,8 @@ class OpenAIProvider extends LLMProvider {
       stream,
       reasoning,
     };
-    const chat = await this.chat(chatRequest);
-    return chat.message.content;
+    const {result} = await this.chat(chatRequest);
+    return result.message.content;
   }
 
   async embed(embedRequest) {

package/lib/llm/utils.js

@@ -379,6 +379,27 @@ function splitThinkTags(input) {
   };
 }
 
+function getContextUsage(usage, numCtx = 65536, maxTokens = 4096) {
+  const used = usage?.prompt_tokens ?? usage?.prompt_eval_count;
+
+  if (!Number.isFinite(used)) {
+    return null;
+  }
+
+  const available = Math.max(0, numCtx - used);
+  const reserved = Math.min(maxTokens, available);
+  const remaining = available - reserved;
+
+  return {
+    used,
+    reserved,
+    remaining,
+    total: numCtx,
+    percentage: Math.min(100, (used / numCtx) * 100),
+    nearLimit: used + maxTokens >= numCtx,
+  };
+}
+
 module.exports = {
   chunkArray,
   embedChunksInBatch,
@@ -394,4 +415,5 @@ module.exports = {
   groupSentences,
   buildPromptFromArticlesMulti,
   splitThinkTags,
+  getContextUsage,
 };

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "goblin-agents",
-  "version": "3.1.0",
+  "version": "3.2.0",
   "description": "AI Agents framework",
   "author": "Epsitec SA",
   "contributors": [